evals-first

To test prompt logic before implementation, you enforce an eval-first process by writing an evaluation that fails before modifying any prompt logic or skill routing in your LLM-interacting code.

Eval-first discipline is a development process that requires writing and verifying behavioral evaluations for LLM-interacting code before the actual implementation is made, ensuring prompt and skill logic correctness.

Yes, this process provides LLM-behavioral regression coverage specifically designed for code that interacts with Claude Code sessions, complementing traditional unit testing.

You need to write evaluations for prompt logic when a new feature is proposed that affects prompt logic or skill routing, activating this process to verify the evaluation before implementation changes.

Eval-first testing does not replace traditional unit testing; it complements it by providing LLM-behavioral regression coverage to ensure the correctness of prompt and skill logic.

To start prompt testing for skill validation, load the evals-first skill and write an evaluation that intentionally fails before you modify the prompt logic or skill in your code.